GitLab

From RiscOS/Env!15, we no longer use aliases, so there's no need to tell
non-interactive shells to expand them.

Various CI jobs run make on all the components within a project. The
COMPONENT and TARGET were correctly set, but we neglected to change into
the appropriate subdirectory first (where applicable).

These jobs were trying to fetch a build tree from a non-existent superproject
`Products/IOMD32` when they should have referenced `Products/IOMDHAL`.

On the Runner machine, each fork of each project gets its own directory,
which is left in the state which the latest job for whatever pipeline was
most recently run on it. This typically will include a large number of
object and binary files, which are of no use to anyone (anything of interest
will already have been packaged up into an artifact and uploaded to the
main GitLab server). Address this by adding an additional job to the end of
each pipeline, which does a `git clean` (it's worth leaving these in place
to reduce the bandwidth requirement when doing a `git fetch` when a pipeline
is next run for the fork).

The Runner machine also stores cache files for each fork of each project, at
least for jobs that complete fully successfully (and there are an increasing
number of these). The way our pipelines use caches, these are tarballs of
pre-built source trees for each target platform. These take up less space
than the temporary files noted above, but will now become the dominant user
of disc space. To address this, abandon use of GitLab Runner's own cache
facility, and take advantage of the fact that shell executors actually have
visibility of the gitlab-runner user's whole home directory to maintain a
single, cached version of each tarball, shared across all forks of all
projects. This is stored within ~/cache, but namespaced under ~/cache/common
to avoid collisions with any users of GitLab Runner's cache facility.

Once the following have been merged:
* Products/BCM2835!6
* Products/BuildHost!2
* Products/Disc!6
* Products/iMx6!2
* Products/IOMDHAL!3
* Products/OMAP3!3
* Products/OMAP4!3
* Products/OMAP5!3
* Products/Titanium!3
* Products/Tungsten!3
then we can have the pipelines for submodules fetch their source trees from
the central projects rather than my forks.

For these versions, "git submodule update --remote" will fail for any
submodules that don't specify a branch in .gitmodules, and which are
currently checked out on a remote tracking branch that tracks a remote
other than "origin". For our pipelines, this means almost any submodule
which has a CrossCompilationSupport branch which has not yet been merged.

Work around it by first doing "git submodule update" without "--remote",
which will put all submodules into detacehd HEAD state.

Resolves bug introduced in commit 9c61e2f0.

If the following happens:
* a GitLab runner recursively clones a superproject
* one or more submodules has changes upstream and the submodule references in
  the superproject have been updated to point to it
* a new pipeline is launched for the new superproject revision

then, while the runner would fetch changes to the superproject, it wasn't
doing so for the submodules. Ironically, we do both `git submodule update`
and `git submodule update --remote` in different pipeline stages and the
latter includes an implicit submodule fetch - but that was the later stage
which we don't get as far as.

To fix this, while retaining the pipeline stage order, include an explicit
fetch in the earlier stage (and remove the implicit fetch in the later one,
since that now merely wastes time).

This one is a special case where the superproject name mismatches the
corresponding Env (and thus Components) file.

Our own CI script also didn't previously support the *removal* of an
autogenerated YAML file. This should be a rare occurrence, but it's best
to automate this also. Also support force-pushing to the submodule; this
will be required in most cases where the pipeline was triggered by a
force-push to *this* project.